Rebuilding comeNpense: From Concept to Live App

comeNpense went through three chapters. First, a collaborative build with a developer partner, where I owned the product side end to end: research, user stories, prioritisation, roadmap. That build paused when my developer’s availability ran out.

Rather than wait, I picked it back up solo using a no-code tool, running my own two-week sprint with daily check-ins and a proper retrospective. That version taught me a lot, but the free plan I was using turned out to have hard limits: I couldn’t publish it anywhere for people to actually use, and some of the features I’d planned (like uploading a bank statement and getting instant results) weren’t possible within those limits either.

So I went back to the drawing board a third time. I kept everything that was already working: the brand, the visual identity, the core promise (know your yearly income and expenses in under 30 seconds, and share that summary in one click). I cut what had made the previous versions heavy or hard to finish, like automatically reading uploaded bank documents, and replaced it with a faster, simpler way to add and import your numbers. I rebuilt the whole thing using an AI-assisted app builder, which let me go from a written product plan to a live, working app in one sitting, on my own.

comeNpense is live today. It’s a small case study in a bigger lesson I keep relearning: a stalled project isn’t a failed one. Sometimes it just needs a different tool and one more pass at deciding what actually matters.

Simplicity requires ruthless prioritisation. Every feature removed is a conscious decision, not a defeat.

Time-boxed delivery only works with full team alignment. Misalignment is a product risk, not just a people problem.

Even 3–5 user conversations change decisions. Research doesn’t need to be large to be valuable.

Building solo across all disciplines is exhausting and illuminating. It deepened my respect for developers and designers enormously.

Where AI could add value: auto-categorising imported transactions, OCR extraction from photographed receipts and payslips, and a simple year-end forecast (“what you’re on track to owe or save”).

What data it would need: transaction-level data (merchant, amount, date, category), a labelled set of categories to ground the classifier, and receipt images for OCR.

What I deliberately scoped out, and why: OCR, bank integrations, forecasting, and any tax engine. The rebuild’s whole point was shipping the core promise fast and reliably, knowing your numbers in under 30 seconds and sharing them in one click. Auto-extraction adds real accuracy risk, a mis-read receipt, a wrong category, before I’ve proven anyone wants the core loop at all. Manual and CSV entry are slower, but they’re honest: what you see is what you entered, nothing quietly guessed on your behalf.

Live since 12 July 2026 · try comeNpense

I’m open to discussing permanent PO/PM roles. Let’s talk about what you’re building.